Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation Antonio Delgado-Rosa a , David Muñoz-Valero a , Enrique Adrian Villarrubia-Martin b , Juan Moreno-Garcia a,∗
arXiv:2607.28470v1 [cs.AI] 30 Jul 2026
a
Escuela de Ingeniería Industrial y Aeroespacial de Toledo, Department of Technologies and Information Systems, Universidad de Castilla–La Mancha, Avenida Carlos III, s/n, Toledo, 45071, Spain b Escuela Superior de Informática, Department of Technologies and Information Systems, Universidad de Castilla-La Mancha, Paseo de la Universidad 4, Ciudad Real, 13071, Spain
Abstract Airborne surveillance from low Earth orbit is hindered by two interconnected bottlenecks: nanosatellites have a limited downlink budget, yet the conventional approach still transmits terabytes of raw imagery to the ground for processing, and open satellite datasets for aircraft are scarce and severely class-imbalanced. These limitations either delay timely decision-making or prevent standard detectors from learning robust representations of rare aircraft classes. In this paper, a workflow that combines on-board inference with generative data augmentation is proposed to address both limitations jointly. Inference is executed on a 6U CubeSat equipped with a low-power edge tensor accelerator, while a diffusion model fine-tuned through low-rank adaptation generates synthetic minority-class imagery. This synthetic output is automatically annotated, pseudo-labelled, by an intermediate detector and merged with classically augmented samples. The results show that the balanced dataset increases global mean average precision from 77.9% to 82.2%, with the minority class rising from F1 = 0.683 to F1 = 0.811, and that the ∗
Corresponding author Email addresses: [email protected] (Antonio Delgado-Rosa ), [email protected] (David Muñoz-Valero ), [email protected] (Enrique Adrian Villarrubia-Martin ), [email protected] (Juan Moreno-Garcia
)
quantised detector fits the on-chip memory and projects 25–30 frames per second on orbit. This approach contrasts with the conventional bent-pipe architecture, in which the satellite acts as a passive data collector. Therefore, the computational tests support the proposed workflow as a decision-support tool for real-time, autonomous airborne surveillance from nanosatellites. Keywords: Satellite Aircraft Surveillance, Edge Computing, Small Object Detection, Data Augmentation, LoRA 1. Introduction Continuous monitoring of the airspace and the Earth’s surface is essential for safety, emergency management and air-traffic control. Traditionally, these surveillance tasks have relied on ground-based infrastructure, such as primary and secondary radar networks, or on the manual review of aerial imagery. However, these approaches share well-known limitations: incomplete geographical coverage, high dependence on expensive installations, and response times that may delay critical decision-making [1, 2]. Over the last decade, the irruption of nanosatellites has democratised orbital observation. Their reduced size and manufacturing cost have enabled the deployment of constellations that offer a high spatial resolution together with a revisit frequency never reached before [3, 4]. Despite these operational advantages, the massive acquisition capacity of on-orbit sensors has given rise to a new technological bottleneck: the volume of imagery generated far exceeds the capacity of traditional radio-frequency links to download it to the ground [5]. The challenge is to process this immense quantity of visual telemetry efficiently, in order to detect relevant targets without requiring constant human intervention. This work is motivated by two complementary gaps that, to the best of the authors’ knowledge, have not yet been addressed jointly in the literature. The first gap concerns the architectural paradigm of current Earth-observation missions, which still follows a centralised conception inherited from the early space era, commonly denoted as bent pipe. Under this scheme, the satellite acts as a mere relay in space: the entire payload of raw imagery is downlinked, and the analysis is delegated to high-performance ground servers. While these servers are able to execute dense predictive models, they relegate advanced orbital systems to the passive role of data collectors. As a result, on-board subsystems lack the situational consistency required to au2
tonomously deduce which fragments of the gathered telemetry possess real operational value. The second gap concerns the training data: open satellite datasets for aircraft are scarce and severely imbalanced. In the standard HRPlanesV2 [6] dataset, military aircraft outnumber the minority helicopter class by a factor of four, which prevents standard one-stage detectors from learning robust representations of rare classes. The state of the art usually addresses these two gaps in isolation: either algorithms are run on the ground with the latency penalty of the downlink, or theoretical models are trained on orbit-ready data that suffocate minority classes, or synthetic images are produced without considering the thermal and power constraints of space hardware. This paper unifies all these dimensions into a single coherent workflow. To overcome the first gap, the workflow executes inference on board a CubeSat of the 6U form factor equipped with a Google Coral Edge Tensor Processing Unit (TPU) as the On-Board Data Processing (OBDP) accelerator. The detector is constrained to fit the 8 MB of on-die Static RandomAccess Memory (SRAM) available on the accelerator after INT8 quantisation, an explicit Size, Weight, and Power (SWaP)-aware design condition that drives both the architecture selection and the compression stage. To overcome the second gap, a generative data-augmentation pipeline based on FLUX [7] and Low-Rank Adaptation (LoRA) [8] is proposed: a custom LoRA token is trained on the minority class, the model is quantised in GGUF format to keep the fine-tuning within reach of consumer hardware, and the synthesised images are pseudo-labelled by an intermediate detector and merged with classically augmented samples. Classical augmentation alone is shown to be insufficient for the helicopter class (F1 = 0.683), whereas the addition of synthetic data lifts it to F1 = 0.811, approaching the performance of the majority classes. The main contributions of this paper are the following: • A hardware-first methodology for space-grade AI deployment. The physical constraints of the target platform, namely the 6U CubeSat bus, the Google Coral Edge TPU, and the 8 MB SRAM budget, are established first, and only then is a detector architecture selected and adapted to fit within them. • A generative data-augmentation pipeline targeted at the minority class. A custom FLUX+LoRA workflow produces synthetic 3
helicopter imagery in four operational environments, which is afterwards pseudo-labelled by an intermediate YOLO detector and merged with classically augmented samples to yield a balanced dataset. • A SWaP-aware benchmark of one-stage detectors for spacegrade hardware. Three architectures (SSD MobileNet V3, YOLO11n and RT-DETR-L) are compared under the same unbalanced aerial dataset and against the same 8 MB SRAM budget of the Coral Edge TPU. YOLO11n emerges as the best precision–latency compromise and is kept as reference. • An on-board concept of operations on a 6U CubeSat. The selected detector is exported to INT8, mapped to the Edge TPU SRAM, and integrated with Slicing Aided Hyper Inference (SAHI) evaluation of 1024×1024 patches, achieving 82.2% mean Average Precision (mAP@50) with a projected on-orbit throughput of 25–30 Frames Per Second (FPS). The source code, FLUX LoRA adapters, ComfyUI workflows and trained weights are released in an open repository (https:// github.com/Antonio23013/TFG-DETECCION-DE-OBJETOS) to ensure that all experiments can be fully reproduced and expanded upon by the community. The rest of the paper is structured as follows: Section 2 reviews recent works showing the progress made in object detection in aerial imagery, generative data augmentation and edge computing for space. Due to the scarcity and strict physical limitations of space-grade components, this work adopts a hardware-first methodology: the physical capabilities must be established before an artificial intelligence model can be feasibly selected. Consequently, Section 3 describes the mission and platform constraints of the 6U CubeSat, including the optical payload, the Edge TPU accelerator, and the communications subsystem. Following this hardware definition, Section 4 presents the materials and methods, namely the HRPlanesV2 dataset, the dataaugmentation pipeline, and the candidate architectures evaluated against those physical constraints. Section 5 describes the experiments and results obtained, showing the performance of the selected architecture in the unbalanced, classically augmented, and synthetically balanced scenarios, and analysing the deployable INT8 model on orbit. Section 6 discusses the practical applicability of the proposal in Intelligence, Surveillance, and Reconnaissance (ISR) and Geospatial Intelligence (GEOINT) scenarios, while acknowl4
edging its limitations. Finally, Section 7 sets out conclusions and further work derived from this paper. 2. Related Work This section reviews the literature most relevant to the objective presented in this work. The review is organised around the three technological threads that the proposed workflow integrates: edge computing for space missions, generative data augmentation, and object detection in aerial imagery. 2.1. Edge computing for space The state of the art in satellite observation reveals a significant technological gap in the communications architecture. The prevailing operational paradigm follows a centralised conception inherited from the early space era, denoted as bent pipe, in which the satellite acts as a mere relay and the entire raw payload is downlinked for ground processing. Under this scheme, on-board subsystems lack the situational awareness required to autonomously discard redundant telemetry, which perpetuates the downlink bottleneck discussed in Section 1. Moving the On-Board Data Processing (OBDP) on board transforms the platform from a passive sensor into an autonomous node, but it demands accelerators compatible with the Size, Weight and Power (SWAP) budget of a Cube Satellite (CubeSat) and, at the same time, resilient to the radiation environment of Low Earth Orbit (LEO). Two reference missions have pioneered this transition: the European Space Agency European Space Agency (ESA) φ-sat-1, which demonstrated on-board cloud screening with a Myriad 2 vision processor, and the ESA CISERES mission [9], led by Deimos, which applies on-board Artificial Intelligence (AI) for civil-security applications. A growing body of work has characterised the suitability of commercial off-the-shelf accelerators for such missions, including Intel Movidius Myriad, the Google Coral Edge Tensor Processing Unit (TPU) and the NVIDIA Jetson family [10, 11, 12], as well as their behaviour under radiation and single-event effects [5, 13, 14]. In the same line, Magalhães et al. [15] benchmarked several edge devices for single-stage object detection in this journal, reporting that the trade-off between precision, latency and power is highly device-dependent. The present work adopts the same benchmarking philosophy but targets a different domain and an additional constraint: the selected detector must fit the on-die 5
Static Random Access Memory (SRAM) of the accelerator after 8-bit Integer Quantisation (INT8) quantisation, so that the whole inference path can run without external memory and within the power envelope of a 6U CubeSat. 2.2. Generative data augmentation Deep-learning detectors remain strongly dependent on the quality and volume of training data. In the aerospace domain, manually collecting and annotating tens of thousands of captures of specific aircraft — such as military helicopters in varied operational environments — is logistically unfeasible, which motivates the use of synthetic data. Generative Adversarial Networks (GANs) [16] were the first generative family applied to this end, with variants such as CycleGAN used to simulate infrastructure or to alter meteorological conditions in aerial photographs. However, GANs suffer from training instability and difficulty in preserving the strict geometric coherence of an aircraft, a limitation that has been extensively documented in surveys on data augmentation for deep learning [17] and in studies on the background bias that arises when synthetic samples are not sufficiently diverse [18]. The recent breakthrough in this area is the use of Latent Diffusion Models (LDMs) [19], which produce images by iteratively denoising a latent representation and achieve clearly increased photometric fidelity to their predecessors. Adapting a diffusion model with billions of parameters to a specific aircraft class would normally require supercomputing resources, but the LowRank Adaptation (LoRA) technique [8] makes this feasible by freezing the pretrained weights and injecting low-rank trainable matrices, which reduces the trainable parameter count by more than 99 %. The same generative principle has recently been applied to minority-class augmentation: Mueller et al. [20] proposed an attention-enhanced conditional-diffusion model for data synthesis in machine fault diagnosis, reporting consistent gains for the under-represented classes. The workflow proposed in the present paper transports this idea to the aerospace domain, fine-tuning a state-of-the-art diffusion model with a custom LoRA token on the minority helicopter class and pseudo-labelling the synthetic output with an intermediate detector, an integration that, to the authors’ knowledge, has not been previously reported. 2.3. Object detection in aerial imagery Before the popularisation of neural networks, aircraft detection in aerial imagery relied on hand-crafted features, exemplified by the combination of 6
Histogram of Oriented Gradients (HOG) descriptors [21] with Support Vector Machine (SVM) classifiers [22]. These early models sought geometric silhouettes, projected shadows or sharp edges of aircraft on runways, but they were extremely rigid: they failed under illumination changes, partial occlusion by clouds and variations in the acquisition angle, and produced high false-positive rates against infrastructure of similar shape. The introduction of Convolutional Neural Networks (CNNs) allowed the network itself to learn geometric and textural features during training, marking the transition towards two-stage detectors such as Region-based Convolutional Neural Network (R-CNN) and, later, Faster R-CNN [23], in which a Region Proposal Network (RPN) first proposes candidate regions and a classifier then confirms the presence of the target. The publication of the Dataset for Object Detection in Aerial Images (DOTA) dataset [24] consolidated this line by providing hundreds of thousands of annotated aerial instances with Faster R-CNN as a baseline. Although two-stage detectors achieved unprecedented precision, their bifurcated architecture was computationally expensive: while one-stage detectors such as You Only Look Once (YOLO) [25] reached more than 30 Frames Per Second (FPS), Faster R-CNN typically remained below 10 FPS, hindering its use in embedded real-time applications. To mitigate the extreme imbalance between background and small targets that characterises aerial scenes, Lin et al. [26] introduced the Focal Loss in RetinaNet, a contribution that directly motivated the present work, since the small object detection problem is central to aircraft recognition from LEO. Modern YOLO versions have since become the industrial standard for aerial imagery, and a number of recent studies in this very journal have explored their behaviour for aircraft detection: Liu et al. [27] proposed a corner-clustering approach coupled with deep learning for aircraft detection in remote-sensing imagery, İlmak et al. [28] assessed YOLOv8 and v9 for efficient plane detection in very high resolution imagery, and Liu et al. [29] provided a comprehensive review of YOLO-based detection in remote sensing. Regarding the smallobject regime specifically, Zhang et al. [30] introduced an edge-aware neural network that explicitly addresses the low signal-to-background ratio of aerial targets. The present work builds on this trajectory but constrains the design from the outset by the SWAP budget of a nanosatellite, a condition absent from the aforementioned studies.
7
2.4. Synthesis and research gap In summary, the reviewed literature shows that one-stage detectors, particularly modern YOLO variants, have become the de facto standard for aircraft detection in aerial imagery, that generative data augmentation based on diffusion models and low-rank adaptation has emerged as a viable alternative to classical augmentation for minority classes, and that on-board edge accelerators have reached a maturity sufficient for satellite deployment. However, these three threads have evolved in isolation: no previous work integrates SWAP-aware architecture selection, generative data augmentation of the minority class and on-board INT8 deployment into a single workflow for airborne surveillance from nanosatellites. The present paper addresses precisely that gap and validates the resulting pipeline on a public aerial dataset with full source-code reproducibility. 3. Mission and Platform Constraints This section sets out the physical and operational context that constrains the rest of the workflow. The nanosatellite platform is presented first, together with the data-volume bottleneck that motivates on-board processing; the on-board AI accelerator and its memory constraint are then characterised; and finally the communications subsystem and the concept of operations that close the loop with the ground segment are described. The architecture selection in Section 4 and the deployment analysis in Section 5 are both driven by the constraints collected here. 3.1. CubeSat form factor and orbital window Historically, space observation has been dominated by monolithic satellites with development cycles often exceeding a decade, multi-million-dollar budgets and operational masses of several tonnes [31]. Over the last two decades, however, the constant miniaturisation of electronics, the adoption of Commercial Off-The-Shelf (COTS) components and the progressive cost reduction of launch systems have driven a genuine revolution in space architecture [3]. Within this transition, the New Space paradigm has categorised small satellites by mass, and nanosatellites platforms between 1 and 10 kg now occupy a fundamental niche for emerging orbital services [32]. The CubeSat standard has been central to this evolution: its predictable geometric units simplify the design of internal subsystems and enable deployment through standardised dispensers such as the P-POD [33], which allows 8
nanosatellites to ride to orbit as secondary payloads and reduces launch costs by several orders of magnitude [34]. In this work, a 6U CubeSat, approximately 10 × 20 × 30 cm, ∼8 kg, Table 1, is adopted as the reference platform. Although the AI accelerator itself has a minimal footprint, the 6U format is justified by the optical payload: training the reference detector at a native resolution of 1024 × 1024 pixels demands high-resolution imagery, and the 6U chassis provides the volume required to host the optics and a sufficiently large solar-panel area for the power budget. The platform is assumed to operate in a LEO at roughly 500 km altitude, where the satellite travels at approximately 7.5 km/s and the visibility window over a given ground station is limited to 10–15 minutes per pass [35]. Table 1: Dimensional and mass specifications of the CubeSat standard.
Format 1U 2U 3U 6U
Dimensions (cm)
Max. mass (kg)
10 × 10 × 10 10 × 10 × 20 10 × 10 × 30 10 × 20 × 30
∼1.33 ∼2.66 ∼4.00 ∼8.00
3.2. Data-volume bottleneck and the case for on-board processing The operational success of LEO constellations has triggered a new transmission challenge. Miniaturised high-resolution optical and infrared sensors generate an unprecedented volume of information, and the capacity to acquire photographic telemetry on orbit now exceeds by far the capacity of the downlink to deliver it to the ground [35]. This is not an arbitrary design problem but a barrier imposed by electromagnetics and by the SWAP budget of the platform: a 3U–6U CubeSat typically generates only 10–30 W through its solar panels [4], which restricts the transmit power of the radiofrequency amplifier, and the reduced volume prevents the installation of highgain parabolic antennas. As a result, the mission must rely on low-profile patch antennas [36, 37], whose throughput although sufficient to transmit metadata in real time — remains inadequate to sustain a continuous dump of raw high-resolution frames. This physical limitation consolidates the central proposal of this work: by integrating a one-stage detector on the on-board processor, the downlink 9
ceases to be a funnel saturated by gigabytes of pixels and is used efficiently to transmit only lightweight metadata vectors target coordinates, confidence level and timestamp — of a few kilobytes. The contrast with the conventional bent-pipe architecture, in which the satellite acts as a passive collector, is summarised in Table 2, on-board inference turns a decision latency of hours or days into fractions of a second and frees the downlink for tactically relevant information only. Table 2: Operational comparison between the conventional bent-pipe architecture and the on-board edge processing proposed in this work.
Operational parameter
Conventional (bent pipe)
Proposed (edge)
None (passive capture and local storage)
Advanced (active inference with deep learning)
Transmitted data
Raw full images (gigabytes)
Metadata vectors and detection alerts (kilobytes)
Bandwidth usage
Total saturation of the downlink.
Minimal, avoiding transceiver bottlenecks
Decision latency
Hours or days (ground postprocessing)
Fractions of a second (nearreal-time on orbit)
Energy impact
Sustained high consumption of Optimised, the TPU cost is offthe RF amplifier set by massive RF savings
On-board cessing
pro-
3.3. On-board AI accelerator and memory constraint Moving inference on board requires an accelerator compatible with the (SWaP) budget of a CubeSat, while remaining resilient to the radiation environment of Low Earth Orbit (LEO). Data-centre Graphics Processing Units (GPUs), such as the NVIDIA T4 used for the training experiments in this work, consume a peak power of 70 W [38]. This thermal and electrical footprint is fundamentally incompatible with the total 10–30 W power budget typically available on a standard nanosatellite platform [4]. The Commercial Off-The-Shelf (COTS) micro-accelerator landscape evaluated for space-borne neural inference comprises three main candidates, summarised in Table 3, the Intel Movidius Vision Processing Unit (VPU) (Myriad 2/X), the Google Coral Edge Tensor Processing Unit (TPU), and the NVIDIA Jetson family 10
(Nano/TX2i/Orin). The first two have already been qualified or are under qualification by the European Space Agency (ESA): the Myriad 2 was demonstrated on orbit by the φ-sat-1 mission for cloud screening [39], and the Coral Edge TPU is currently being assessed for radiation tolerance (Total Ionizing Dose and Single-Event Effects) by the ESA CAIRS21 project [14, 12]. Table 3: Technical comparison of the COTS micro-accelerators evaluated for on-board inference.
Platform / chip
Architecture Power
Optimal precision
Space validation
Intel Movidius (Myriad 2/X)
VPU (vision)
∼1.5–2 W
FP16 / INT8
φ-sat-1 2020)
(ESA,
Google Coral (Edge TPU)
ASIC sor)
∼2 W
INT8 (quantised)
CAIRS21 2024)
(ESA,
NVIDIA Jetson (Nano/TX2i)
SoC (GPU, 5–15 W CUDA)
FP32 / FP16
Aitech Venus space computers
(ten-
The Coral Edge TPU is selected as the reference accelerator for the remainder of this work. It delivers up to 4 Tera-Operations Per Second (TOPS) at 2 W (2 TOPS/W) [11], and its compilation ecosystem provides a robust cross-compilation flow that maps a model trained on a general-purpose GPU to the on-board Application-Specific Integrated Circuit (ASIC) [40]: the network is first exported and quantised to INT8 into a .tflite file [41, 42], and the Edge TPU Compiler then assigns the matrix operations directly to the chip instructions. Two strict physical constraints imposed by this device drive the architecture selection in Section 5: • INT8 quantisation. The accelerator lacks the silicon to process 32bit floating-point (FP32) arithmetic; the entire model must be quantised to 8-bit integers. This step reduces the parametric weight by approximately 75 % with only a marginal impact on overall detection precision [42]. • 8 MB on-die SRAM limit. The Edge TPU integrates a fast but physically constrained Static Random-Access Memory (SRAM) of 8 MB directly on the same die as the processor [10]. For optimal performance, the quantised model’s parameters and execution graph must fit entirely within this budget, allowing the full inference path to run in-cache at 11
maximum speed. If the model exceeds this limit, blocks of computation must be offloaded to the general-purpose host CPU of the CubeSat. This offloading triggers massive data transfers across the system bus, causing a severe drop in Frames Per Second (FPS) and negating the energy benefits of the accelerator [40]. The combination of these two constraints defines the central design rule of this paper, the selected detector must, once quantised to INT8, fit strictly within the 8 MB SRAM of the Edge TPU to guarantee real-time, energyefficient performance on orbit. 3.4. Optical payload and attitude control Image acquisition is delegated to COTS optics specific to the New Space sector. Modular optical systems such as the xScape200 series by Simera Sense [43] or the Mantis sensor by Dragonfly Aerospace [44] integrate organically into a 6U volume and provide a Ground Sample Distance (GSD) of approximately 1.5 m/pixel from a 500 km orbit. At this resolution, a standard-sized aircraft occupies a sufficient pixel area on the focal plane to preserve its morphology, which justifies the native input resolution adopted for the detector. To guarantee orthogonal framing of the ground targets, the satellite relies on a three-axis Attitude Determination and Control System (ADCS) combining Star Trackers and Reaction Wheels, maintaining a strict nadir-pointing attitude during the acquisition phase and eliminating geometric distortions due to oblique perspective [45]. 3.5. Communications and concept of operations Because the on-board neural processor autonomously discards empty captures, the volume of data to be transmitted is drastically reduced, the CubeSat only downlinks a lightweight metadata packet, coordinates, timestamp and confidence, together with the compressed 1024 × 1024-pixel patch corresponding to a positive detection. A split communications architecture is therefore proposed, illustrated in Figure 1. • Payload downlink. A COTS S-band transmitter (e.g. EnduroSat SBand Transmitter [46]) operating in the 2.2–2.3 GHz band provides up to 10 Mbps, which is more than sufficient to deliver the detection outputs during the few minutes of visibility over the ground station. A low-profile patch antenna is used so that the radiating element does not interfere with the solar-panel area. 12
• Telemetry, Tracking and Command (TT&C). A standard omnidirectional UHF transceiver is retained for command uplink and healthmonitoring telemetry [47]. The mission operates under a total-decision autonomy model that alternates between low-power and tactical-intervention modes [48]: • Idle and planning. The satellite orbits with the optical and computational payload switched off. It activates only when approaching pre-programmed geographic coordinates of strategic interest, located through the on-board Global Navigation Satellite System (GNSS) receiver. • Dynamic acquisition with circular buffer. While overflying the target zone, the camera does not store captures permanently; the highresolution data stream is buffered temporarily in the Random Access Memory (RAM) of the main on-board computer. • Intelligent transmission. The AI accelerator analyses the volatile buffer in real time. Frames without detections above the confidence threshold are discarded and overwritten immediately; upon a positive detection, the system interrupts the erase cycle, extracts the useful frame, compresses it together with its metadata, and stores it permanently for downlink. This methodology ensures that the space segment delivers exclusively tactical intelligence, maximising the operational autonomy and avoiding the bandwidth restrictions that would otherwise saturate the downlink. The constraints collected in this section 6U form factor, INT8 quantisation and the 8 MB SRAM budget of the Edge TPU, nadir-pointing optics at ∼1.5 m GSD, and the S-band/UHF communications split — are the inputs that drive the methodology and the experimental campaign presented next. 4. Materials and Methods This section describes the materials and methods that support the experimental campaign. The baseline dataset and its initial diagnosis are presented first (Section 4.1); the two-stage data-augmentation pipeline — classical transformations followed by generative synthesis — is then detailed (Section 4.2); and finally, in Section 4.3, the candidate detection architectures 13
SPACE SEGMENT (6U CubeSat in LEO Orbit) Patch Antenna (S-Band) Payload: 2 - 10 Mbps
UHF/VHF Antenna Telemetry & Commands: 9.6 kbps
Link A (High Speed) Visual inference download [e.g., Crop 1024x1024, HELI 85%]
Link B (Low Speed) Telemetry and Positioning [e.g., Coords X,Y / Platform Status]
GROUND STATION Parabolic Antenna (S-Band Receiver)
Yagi/Dipole Antenna (UHF/VHF Receiver)
Receiver and Demodulator
Control Center Tactical Alert Confirmed! (Helicopter at [X,Y], Battery OK)
Figure 1: Block diagram of the space-to-ground data-downlink architecture for the 6U CubeSat. The schematic illustrates the dual-band communication strategy: a high-speed S-band link (2–10 Mbps) for visual inference downloads and a low-speed UHF/VHF link (9.6 kbps) for platform telemetry and tactical alerts.
14
and the evaluation metrics are introduced, together with the SWAP-aware selection criterion that links methodology to the platform constraints of Section 3. 4.1. Dataset: HRPlanesV2 The starting point of the study is the public HRPlanesV2 dataset, hosted on the Roboflow computer-vision platform [6]. It provides a base of highresolution satellite images with structured annotations of aircraft in three operational categories: civil aircraft, military aircraft, and helicopter. The dataset is partitioned into training, validation, and testing subsets, comprising 79 %, 13 %, and 8 % of the total images, respectively. Before any training, the original training partition was diagnosed along two complementary axes that condition the rest of the workflow. The first axis is the quantitative distribution of instances. A total of 30874 valid bounding boxes are available, distributed as follows (see Figure 2): military aircraft is the majority class with 17814 instances (57.7 %), civil aircraft accounts for 8603 instances (27.9 %), and helicopters provide only 4457 instances (14.4 %). The ratio between the majority and the minority class is approximately 4:1. Training a one-stage detector directly on this distribution biases the network towards the majority class when minimising the global loss, which is unacceptable for an early-warning mission and motivates the oversampling strategy described next. The second axis is morphological: the relative area of each bounding box with respect to the full image. A custom Python script was written to parse every annotation file, extract the bounding-box coordinates and normalise the geometric area by the image resolution. The resulting box plot is shown in Figure 3. The analysis illustrates a classical small object detection problem: 62.2 % of civil aircraft, 48.4 % of helicopters and 26.8 % of military aircraft occupy less than 1 % of the total image area. The helicopter class is the most vulnerable, combining the smallest number of samples with the smallest mean area (1.08 %), against 1.893 % for military aircraft. This indicates that simple whole-image oversampling is insufficient and that scalealtering transformations (zoom-in, centred crops) must be included in the augmentation pipeline to force the network to extract low-level features. 4.2. Data-augmentation pipeline To preserve methodological integrity and avoid data leakage, every transformation described below is applied exclusively to the training partition; 15
Class Distribution in Original Dataset (Train)
17814
Number of Instances (Bounding Boxes)
17500 15000 12500 10000
8603
7500 5000
4457
2500 0
Civil Aircraft
Helicopter
Operational Categories
Military Aircraft
Figure 2: Distribution of instances (bounding boxes) per category in the original training set of HRPlanesV2.
Relative Area Distribution of Aircraft
Bounding Box Area (% of image)
4
3
2
1
0 Civil Aircraft
Helicopter
Operational Categories
Military Aircraft
Figure 3: Distribution of the relative area of the bounding boxes with respect to the total image size, per operational category. Outliers are omitted for readability.
16
the validation and test partitions are left untouched so that the final metrics reflect performance on unaltered real scenarios. 4.2.1. Classical augmentation The first stage applies classical transformations to the original pixels to systematically increase morphological and radiometric diversity without requiring additional data collection. Three distinct transformation families are combined, all implemented through the widely adopted Albumentations library [49]: • Geometric transformations alter the spatial arrangement of pixels without modifying their intensity. Horizontal and vertical flips, random rotations and centred crops/zoom-in operations are applied. The latter directly addresses the small-object problem identified above by forcing the target to occupy a larger fraction of the image. • Photometric transformations modify the intensity values and colour channels: HSV/RGB jittering, Gaussian and salt-and-pepper noise, and Gaussian blur. These emulate the radiometric variability that an on-orbit optical payload faces across passes (thermal noise, solar glint, partial cloud cover, motion blur at 7.5 km/s). • Mixed transformations apply geometric and photometric operations sequentially and probabilistically on the same image, generating complex edge-case samples that prevent premature convergence and reinforce robust feature extraction in the convolutional layers. This stage produced 2832 augmented images (928 geometric, 939 photometric and 965 mixed), bringing the minority classes towards an approximate balance of 18 000 instances per class. 4.2.2. Generative augmentation with FLUX and LoRA Classical augmentation is bounded by the information contained in the original pixels: it can transform existing data but cannot generate new knowledge. If no original image depicts a helicopter overflying a snowfield or a desert, no classical transformation can recreate that scenario. To overcome this representational limit, a generative stage based on Latent Diffusion Models (LDMs) is introduced. Generative Adversarial Networks (GANs) [16] were discarded because of their training instability and difficulty in preserving the strict geometric coherence of an aircraft, as documented in surveys on 17
augmentation for deep learning [17] and in studies on the background bias that arises from insufficiently diverse synthetic samples [18]. The selected generative backbone is FLUX [7], which replaces traditional diffusion with flow-matching and provides the photorealistic fidelity required for military fuselages and helicopter rotors. Adapting a billion-parameter diffusion model to a specific aircraft class would normally demand supercomputing resources. This is made feasible by Low-Rank Adaptation (LoRA) [8], which freezes the pretrained weight matrix W0 and approximates the update ∆W as the product of two lowrank matrices B and A: Wnew = W0 + B A, (1) where, if W0 ∈ Rd×k , then B ∈ Rd×r and A ∈ Rr×k with intrinsic rank r ≪ min(d, k). For a 10 000 × 10 000 matrix (100 M parameters), the decomposition with r = 8 reduces the trainable count to 160 000 (−99.8 %), which makes fine-tuning tractable on consumer hardware. The concrete pipeline operates as follows (Figure 4). The base model FLUX.1-dev is loaded in quantised GGUF format (Q4_0) so that the finetuning fits comfortably within the 16 GB of VRAM available on the university workstation. A custom LoRA adapter is trained on the minority helicopter class with the trigger token sathelo. Inference is run through a ComfyUI workflow composed of a UNET GGUV loader, a dual CLIP loader (CLIP-L and T5-XXL) for text conditioning, the LoRA adapter applied with model strength 0.85 and clip strength 1.0, a KSampler configured with 27 steps, CFG = 1, Euler sampler and Beta scheduler at denoise 0.95, and a VAE decode stage. Synthetic helicopter imagery is generated across four operational environments to maximise contextual diversity: industrial/airport, rural/forest, coastal/maritime and arid/desert, yielding approximately 2 226 helicopter images (see Figure 4). For the civil class, classical Albumentations transforms (flip, HueSaturationValue, RandomBrightnessContrast, GaussNoise, Blur, ShiftScaleRotate) are applied to existing real samples. 4.2.3. Pseudo-labelling and class balancing Manually annotating thousands of synthetic images is unfeasible. An intermediate YOLOv11 detector, previously trained on the classically augmented set, is used to auto-label the synthetic output. The resulting labels are merged with the classically augmented samples, producing a final balanced dataset of more than 53 000 instances at approximately 18 000 instances per class. 18
REAL DATA (Limited helicopter samples)
Adapter training GENERATIVE PIPELINE (FLUX.1 + LoRA) CLIP Text Encode (Environment prompt: snow, desert, etc.)
LoRA Adapter (Token: sathelo)
Checkpoint (FLUX.1 Base Model)
K-Sampler (Diffusion Process)
Generated images SYNTHETIC OUTPUT (Class Balancing)
AUGMENTED DATASET (YOLO11 Training)
Figure 4: Block diagram of the generative data-augmentation pipeline. The limited real helicopter instances are used to train a custom LoRA adapter. During inference, the FLUX.1 base model, combined with the domain-specific LoRA and environmental text prompts, feeds the K-Sampler to generate a diverse synthetic output, balancing the final YOLO11 dataset.
19
4.3. Candidate architectures and evaluation metrics Three one-stage detectors are evaluated as candidates for the on-board mission, chosen to span the precision–efficiency design space: • SSD MobileNet V3 Large [50]: a depthwise-separable backbone that represents the historical standard for ultra-lightweight deployment. • YOLO11n [51, 25]: the nano version of the latest YOLO release, built around a C3K2 backbone with Spatial Pyramid Pooling – Fast (SPPF) and a Path Aggregation Network (PANet) neck, plus a C2PSA spatialattention head. • RT-DETR-L [52, 53]: a hybrid convolutional/Transformer detector that replaces the anchor-based head with an end-to-end bipartite-matching decoder, eliminating Non-Maximum Suppression (NMS). All models are evaluated with the same metric set, which jointly captures detection quality and deployability: mAP@50 and mAP@50-95 for detection quality; Precision, Recall and F1-Score per class for tactical behaviour; and parameters, weight on disk, GFLOPs and FPS for computational cost. The SWAP-aware selection criterion is defined explicitly at this point and applied in Section 5: a candidate is admissible only if, after INT8 quantisation, its parametric weight fits within the 8 MB on-die SRAM of the Coral Edge TPU (Section 3.3); otherwise the inference path falls back to the generalpurpose Central Processing Unit (CPU) and the on-orbit FPS collapses. This criterion is the link between the mission-and-platform constraints of Section 3 and the experimental campaign that follows. 5. Experiments and Results This section details the empirical evaluation of the proposed methodology. We first outline the hardware and software environments used for training, alongside the specific hyperparameters and configuration choices governing the experimental pipeline (Section 5.1). Subsequently, we present a comparative benchmark of the three candidate one-stage architectures— SSD MobileNet V3, RT-DETR-L, and YOLO11n—evaluated on the original, unbalanced HRPlanesV2 dataset to justify the selection of YOLO11n based on its precision, inference speed, and edge-compatibility constraints (Section 5.2). 20
5.1. Experimental Setup and Hyperparameters To ensure full reproducibility of the computational experiments, the training configurations were standardized across all evaluated models. The experiments were conducted in a Google Colab environment utilizing a dedicated GPU accelerator. For the final YOLO11n training on the augmented dataset, the network was initialized with pre-trained COCO weights (yolo11n.pt) to accelerate convergence and subsequently fine-tuned. Table 4 summarizes the primary hyperparameters explicitly configured during the training phase, including the exact random seed used to guarantee identical replication of the reported metrics. Unlisted hyperparameters were kept at their default Ultralytics settings. Table 4: Training hyperparameters for the YOLO11n network.
Hyperparameter
Value
Input Resolution Epochs Batch Size Random Seed Initial Weights
640 × 640 pixels 20 Auto (memory-optimized, batch = -1) 93 yolo11n.pt
Preliminary training was carried out on a Google Colab Graphics Processing Unit (GPU) NVIDIA T4, and the final runs on an NVIDIA RTX 4080 16 GB workstation. To explicitly rule out stochastic optimization bias and characterize the variability of the network, a five-seed study was conducted (seeds 0, 16, 42, 93, and 2026). The model’s performance remains highly stable across initializations, yielding a global mean mAP@50 of 0.805 with negligible variance. Because the architecture is not affected by random weight initialization, seed 93—which closely aligns with the average validation losses and precision-recall trade-off—was systematically selected as the representative model for all subsequent inference and quantization experiments. The full stochastic study is released in the public repository for reproducibility. All models were trained with the AdamW optimiser, an automatic batch size (AutoBatch = -1), Automatic Mixed Precision (AMP) (FP16), and the mosaic-augmentation schedule recommended by Ultralytics for Real-Time Detection TRansformer (RT-DETR). A resolution sweep between 640 and 1024 pixels was performed for YOLO11n to quantify the impact of the input size on small-object detection. 21
Table 5: Comparative benchmark of the three candidate one-stage detectors on the unbalanced HRPlanesV2 dataset. The Edge TPU admits only models whose INT8-quantised weight fits its 8 MB on-die SRAM.
Model SSD MobileNet V3 RT-DETR-L YOLO11n YOLO11n
Resolution
Params
Weight
mAP@50
FPS
320 640 640 1024
3.07 M 32.81 M 2.59 M 2.59 M
11.93 MB 188.85 MB 5.22 MB 5.22 MB
44.98 % 74.53 % 75.01 % 77.90 %
82.66 14.05 48.28 34.46
As detailed in Table 5, the candidate models were evaluated to determine their suitability for on-board edge deployment. SSD MobileNet V3 achieved the highest inference speed (82.66 FPS) but was discarded due to a severe background collapse, yielding an unacceptable mAP@50 of 44.98 %. Conversely, while RT-DETR-L demonstrated strong detection capabilities, it was ruled out because its 188.85 MB memory footprint heavily exceeds the strict 8 MB on-die SRAM limit of the Edge TPU. Ultimately, YOLO11n was selected as the optimal architecture. Both its 640 and 1024-pixel resolution variants comfortably fit within the hardware memory constraints (5.22 MB). The 1024-pixel configuration was chosen as the definitive baseline, as it maximizes the small-object detection performance (77.90 % mAP@50) while maintaining a highly capable real-time inference rate of 34.46 FPS. 5.2. Architecture benchmark (unbalanced dataset) The three candidate architectures were first trained and evaluated on the unbalanced HRPlanesV2 partition described in Section 4.1. The resulting figures are summarised in Table 5 and comprehensively analyzed across three key visual dimensions: the balance between detection capacity and inference speed, the hardware constraints regarding memory and parametric load, and the operational design space. As illustrated in Figure 5, SSD MobileNet V3, despite being the fastest and lightest candidate (3.07 M parameters, 11.93 MB), collapses onto the background class with a mAP@50 of only 44.98 %. This critical failure is due to an incompatibility between its ImageNet-pretrained backbone and the distinct statistical distribution of the aerial scenes. Conversely, RT-DETR-L achieves a competitive 74.53 % mAP@50, but as clearly depicted in Figure 6, its 188.85 MB footprint exceeds the 8 MB SRAM budget by more than an 22
SWaP Constraints: Memory Footprint and Parametric Load 40
Parameters (Millions) Disk Weight (MB)
175
32.81M
30
150
25
125
20
100
15
75
10
50
5 0
3.07M
2.59M
11.93 MB
MobileNet V3
RT-DETR-L
Disk Weight (MB)
Parameters (Millions)
35
200
188.85 MB
25 5.22 MB
YOLO11n
0
Figure 5: Direct comparison between mean average precision (mAP@50) and baseline inference speed.
order of magnitude. Even after INT8 quantisation, the model cannot reside in the accelerator cache, and its attention-specific matrix operations are not supported by the integer Arithmetic Logic Unit (ALU) of the Edge TPU, forcing a fallback to the CPU and consequently collapsing the inference speed to 14.05 FPS. YOLO11n offers the optimal precision–efficiency compromise. Figure 7 visually encapsulates this advantage: it is the only candidate that simultaneously reaches the upper-right (high precision, high speed) region while maintaining a minimal radius, indicative of its negligible memory footprint. At 640 px, it already matches the precision of the Transformer-based detector at more than three times the speed (48.28 vs. 14.05 FPS). Furthermore, increasing the input resolution to 1024 px lifts mAP@50 to 77.90 % and mAP@50-95 by 6.65 percentage points, with a particularly strong impact on the helicopter class (where the acquisition rate rises from 46 % to 57 %). The operating point on the F1 curve is fixed at a confidence threshold of 0.78, where the F1 score reaches 0.78. The residual weakness of the unbalanced YOLO11n model lies in the helicopter class: in the normalised confusion matrix, 34 % of helicopter instances 23
Direct Comparison: Detection Capacity vs. Refresh Rate
Precision mAP@50 (%)
Inference Speed (FPS)
100
82.66
80
80
75.01%
74.53%
60
60 48.28
44.98%
40
40
20
20
14.05
0
MobileNet V3
Inference Speed (FPS)
Precision mAP@50 (%)
100
RT-DETR-L
YOLO11n
0
Figure 6: Parametric footprint of each architecture against the 8 MB SRAM budget of the Edge TPU. RT-DETR-L exceeds the limit by nearly 190 MB; YOLO11n fits comfortably after INT8 quantisation.
Operational Trade-off: Precision vs. Speed (Bubble size indicates memory footprint)
90 RT-DETR-L (188.85 MB)
Precision - mAP@50 (%)
80
YOLO11n (5.22 MB)
70 60 50
MobileNet V3 (11.93 MB)
40 30
0
20
40
60
Inference Speed (FPS)
80
100
Figure 7: Operational design space. The bubble size is proportional to the model weight on disk (MB). YOLO11n is the only candidate in the upper-right region with a minimum radius.
24
are misclassified as background, and a residual 19 % confusion between civil and military aircraft remains. This outlines the open problem that the augmentation campaign of Section 5.3 successfully targets. 5.3. Impact of data augmentation Two experiments are compared against the unbalanced YOLO11n baseline of Section 5.2, both trained for 20 epochs at 1024 px with seed 93. 5.3.1. Experiment 1: classical augmentation Applying the classical pipeline of Section 4.2.1 yields 2 832 additional images and brings the classes to an approximate balance of 18 000 instances each. Global mAP@50 rises from 77.90 % to 81.62 %, and the maximum F1 to 0.77 at a confidence threshold of 0.264. The per-class F1, however, exposes an asymmetry: civil aircraft reaches 0.847 and military aircraft 0.887, but helicopters remain at 0.683, with 31 % of helicopter instances still collapsing onto the background. Classical augmentation is therefore effective for the majority classes but insufficient for the minority class.
Impact of Data Augmentation Strategies on Performance 0.85
Validation mAP@50
0.80 0.75 0.70 0.65 0.60 0
2
4
6
8
10
Epoch
12
Baseline Model (Unbalanced) Classic Augmentation Synthetic Augmentation (Generative) 14 16 18 20
Figure 8: Evolution of mAP@50 across the three training regimes: unbalanced baseline, classical augmentation and FLUX+LoRA generative augmentation.
25
5.3.2. Experiment 2: FLUX + LoRA generative augmentation The overall performance progression and the optimal operating threshold of the model are further detailed in Figure 8 and Figure 9. As illustrated in Figure 8, the evolution of the validation mAP@50 metric across the three training regimes underscores the effectiveness of the proposed methodology. While the transition from the unbalanced baseline to the classical augmentation provided initial improvements, it is the introduction of the FLUX+LoRA generative augmentation that maximizes the global performance. Furthermore, the narrow confidence intervals depicted in the learning curves demonstrate that this synthetic balancing is highly robust against stochastic optimization variance. Ultimately, this approach not only increases the global mAP@50 but also ensures a more equitable feature representation across all aeronautical categories, effectively closing the performance gap between the minority and majority classes.
F1-Confidence Curve
1.0
0.8
F1
0.6
0.4
0.2
0.0 0.0
0.2
0.4
0.6
Confidence
Civil Aircraft 0.82 at 0.392 Helicopter 0.79 at 0.392
0.8
1.0
Military Aircraft 0.75 at 0.133 All Classes 0.79 at 0.394
Figure 9: F1 curve of the final YOLO11n model trained on the FLUX+LoRA-balanced dataset, with a maximum F1 of 0.79 at a confidence threshold of 0.394.
26
Furthermore, the F1-confidence curve (Figure 9) evaluates the harmonic mean of precision and recall across different confidence thresholds. The model reaches a maximum F1 score of 0.79 at a confidence threshold of 0.394. This specific optimal threshold aligns with the previously discussed synthetic domain shift: the model requires a moderate, rather than extremely high, confidence threshold to maximize the detection of challenging or degraded real-world instances without succumbing to excessive background false positives. Together, these metrics empirically validate the generative augmentation strategy as a highly effective approach for object detection in imbalanced datasets. To fully characterize the model’s behaviour, a detailed analysis of the combined confusion matrix (Figure 11) is required, which presents both normalized percentages and absolute metrics simultaneously. The matrix indicates that the model possesses strong detection capabilities, correctly identifying 7931 civil aircraft (74%), 3187 helicopters (82%), and 4264 military aircraft (79%). However, particular attention must be given to background misclassifications. The persistence of false negatives—true aeronautical instances predicted as background, totalling 1213 civil, 557 helicopters, and 550 military instances—can be attributed to the nature of the synthetic dataset expansion. While generative augmentation (FLUX+LoRA) successfully balances the class distribution and provides high-quality samples, it can introduce a subtle domain shift. The neural network learns highly specific synthetic feature representations that may occasionally fall short of capturing the full complexity, varied lighting, or sensor noise of extreme real-world edge cases. Consequently, when faced with highly degraded real-world instances, the model adopts a conservative inference strategy, abstaining from predictions and defaulting to the background class. Conversely, the false positive distribution (background regions predicted as aircraft, comprising 694 civil, 756 helicopters, and 890 military instances) highlights the model’s heightened feature-extraction sensitivity. This phenomenon is largely driven by two factors. First, many of these perceived errors are overlapping bounding boxes on a single aircraft, where the evaluation metric penalizes the duplicate as a background false positive. Second, the dataset contains originally unlabeled instances that human annotators missed. The current detector successfully identifies these unannotated targets, effectively outperforming the original ground-truth annotations in certain scenes. This capability is visually corroborated in subsequent qualitative 27
evaluations (e.g., Figure 10), which show the model detecting valid aircraft that were completely absent from the original dataset labels. Regarding the background class, it is important to note that the intersection representing True Negatives (Background-Background) is left intentionally blank. Unlike standard image classification, object detection frameworks evaluate performance based on discrete bounding box proposals. Consequently, the background comprises a virtually infinite number of spatial locations across the image where no object exists and none is predicted. Since True Negatives are mathematically unquantifiable in this spatial context and do not contribute to standard detection metrics such as Precision, Recall, or mAP, this cell is omitted to prevent the distortion of the overall performance statistics. 5.4. Deployable model: INT8 quantisation and on-orbit inference Once YOLO11n is selected, the model is exported through the Ultralytics export pipeline [41] to a .tflite file and quantised to INT8 with the posttraining integer-quantisation toolchain of TensorFlow [42]. The Edge TPU Compiler then maps the matrix operations directly to the chip instructions. Quantisation reduces the parametric weight from 5.22 MB to approximately 2.5 MB (−75 %), so the entire inference path resides in the 8 MB on-die SRAM and runs without CPU offloading. Scaling the measured T4 latency to the Edge TPU throughput-to-power ratio yields a projected on-orbit rate of 25–30 FPS at 1024 px. To handle full-scene acquisitions without degrading the native GSD, inference is wrapped in a Slicing Aided Hyper Inference (SAHI) scheme [54]: each high-resolution capture is sliced into 1024 × 1024 patches with a 15 % perimeter overlap, the Edge TPU evaluates every patch sequentially, and the validated detection vectors are projected back onto the global coordinate system of the original capture, where NMS merges duplicate predictions in the overlap regions. 5.5. Extended multi-class model and ISR application As a proof of concept of the extensibility of the pipeline, the same YOLO11n architecture was retrained on an extended 22-class dataset comprising specific aircraft models (F-16, F-22, B-52, U-2, C-130, E-3, . . . ) plus the Runway class. Evaluated on a strictly isolated test partition to rule out overfitting, the resulting normalised confusion matrix (Figure 12) shows a diagonal above 28
Ground Truth: 11 | Detected: 12
Ground Truth: 8 | Detected: 9
Ground Truth: 7 | Detected: 8 Ground Truth: 7 | Detected: 8
Ground Truth: 3 | Detected: 5 Ground Truth: 18 | Detected: 19
Figure 10: Qualitative evaluation of apparent false positives. The inference results show scenes where the trained model successfully detects valid aircraft that were entirely missing from the original ground-truth annotations (e.g., predicting 14 objects where the dataset only labeled 13). This supports that a portion of the reported background misclassifications are actual correct detections, highlighting the model’s feature extraction capabilities.
29
100
73.93% (7931)
Civil Aircraft
2.33% (250)
12.43% (1334)
11.31% (1213) 80
0.46% (18)
81.93% (3187)
3.29% (128)
14.32% (557)
Military Aircraft
8.46% (454)
1.88% (101)
79.42% (4264)
10.24% (550)
True Label
Helicopter
29.66% (694)
Background
32.31% (756)
60
40
20
38.03% (890) d ckg rou n Ba
Mil
ita
ry A
irc
raf t
pte r He lico
Civ
il A
irc
raf t
0
Predicted Label Figure 11: Combined confusion matrix of the final balanced model, displaying both normalized percentages and absolute instance counts. The diagonal highlights the detection rates across all targeted aeronautical categories.
95 % for almost every class. This high accuracy illustrates the feature extraction of the augmented dataset rather than mere data memorisation. The only exception is Runway, whose accuracy drops to 37 % due to confusion with linear background structures a limitation that motivates the future migration to instance segmentation discussed in Section 7. An illustrative inference mosaic on an airbase scene is shown in Figure 13, illustrating the operational value of the pipeline for ISR and GEOINT applications: the on-board detector localises and classifies aircraft across the captured area, and only the confirmed metadata would be downlinked. Representative synthetic helicopter samples generated by the Flow Matching Text-to-Image Model (FLUX)+LoRA pipeline are shown in Figure 14: the photorealistic fidelity and the diversity of backgrounds support the em30
A-10 1.00 A-26 1.00 0.94 0.15 Airport 1.00 0.03 B-1 1.00 B-2 1.00 0.01 B-29 0.98 0.01 B-52 0.96 0.03 0.010.08 Boeing 0.98 0.03 C-130 0.02 0.99 C-135 1.00 0.01 C-17 0.96 0.03 C-21 0.01 0.99 0.03 C-5 0.01 0.01 1.00 E-3 1.000.01 F-16 0.01 0.99 0.01 F-22 0.02 1.00 0.01 KC-10 1.00 P-63 0.37 0.60 Runway 1.00 T-43 1.00 T-6 0.01 0.990.03 U-2 0.06 0.01 0.62 background A-10 A-26 Airport B-1 B-2 B-29 B-52 Boeing C-130 C-135 C-17 C-21 C-5 E-3 F-16 F-22 KC-10 P-63 Runway T-43 T-6 U-2 background
Predicted
Confusion Matrix Normalized
1.0
0.8
0.6
0.4
0.2
0.0
True
Figure 12: Normalised confusion matrix of the extended 22-class YOLO11n model. The diagonal exceeds 95 % for almost every class except Runway (37 %).
31
0: 19 | 1: 0 | 2: 0 0: 0 | 1: 0 | 2: 35
0: 11 | 1: 3 | 2: 0
0: 8 | 1: 0 | 2: 0
0: 26 | 1: 1 | 2: 0
0: 1 | 1: 1 | 2: 0
0: 41 | 1: 0 | 2: 0
0: 11 | 1: 0 | 2: 0 0: 0 | 1: 8 | 2: 0
LEGEND & TOTAL DETECTIONS
0: Civil Aircraft (117)
1: Helicopter (13)
2: Military Aircraft (35)
Figure 13: Inference mosaic on an airbase scene, illustrating the operational value of the pipeline for Intelligence, Surveillance and Reconnaissance (ISR) and Geospatial Intelligence (GEOINT) applications.
32
pirical gains reported above. Desert Background
Airport Background
Rural Background
Coastal Background
Figure 14: Synthetic helicopter imagery generated by the FLUX+LoRA pipeline across the four operational environments.
6. Discussion This section interprets the empirical findings presented in Section 5, contextualising the performance of the proposed YOLO11n pipeline within the broader framework of edge computing for nanosatellites. Having demonstrated that the combination of a SWaP-aware detector and generative data augmentation successfully overcomes severe class imbalance, we now evaluate the real-world implications of these results. Section 6.1 explores the operational impact of this autonomous architecture for ISR and GEOINT missions, 33
detailing its strategic advantages over traditional bent-pipe paradigms. Finally, Section 6.2 critically examines the boundaries of the current study, identifying methodological constraints and outlining the necessary steps toward physical hardware validation. 6.1. Practical applicability in ISR and GEOINT The experimental results support the proposed workflow as a decisionsupport tool for real-time, autonomous airborne surveillance from nanosatellites. By moving inference on board the 6U CubeSat, the platform evolves from a passive data collector, characteristic of the conventional bent-pipe architecture, into an autonomous node able to filter redundant telemetry and deliver only tactical intelligence. In an ISR or GEOINT context, this translates into two operational advantages. First, the on-orbit inventory of airbases enabled by the extended 22-class model (Section 5.5) allows automated order-of-battle estimation without downlinking raw imagery. Second, the intelligent-transmission ConOps of Section 3.5 reduces the decision latency from hours or days, typical of ground post-processing, to fractions of a second, which is critical for tracking dynamic targets such as departing or manoeuvring aircraft. The trade-off between the conventional and the proposed architectures, quantified in Table 2, shows that the energy spent by the Edge TPU is largely offset by the massive savings in radio-frequency transmission, making the approach energetically sustainable within the 10–30 W budget of a CubeSat. 6.2. Limitations Despite these promising outcomes, the proposed approach still faces a number of limitations that should be addressed before an operational deployment. • Software-only validation. All FPS figures reported in Section 5.4 are throughput projections obtained by scaling the measured T4 latency with the throughput-to-power ratio of the Edge TPU; no Hardwarein-the-Loop test on a real flight-grade accelerator, nor radiation tests (Single Event Effect (SEE)/Total Ionising Dose (TID)), have been conducted yet. The physical deployment of the compiled model on an Edge TPU or a Jetson Orin Nano, with measured power and per-frame latency under vacuum and radiation, is left as future work. 34
• Runway-class performance. In the extended 22-class model, the Runway class drops to 37 % accuracy due to confusion with linear background structures. Bounding-box detection is not the most adequate formulation for elongated infrastructure, and a migration to instance segmentation (e.g. YOLO-seg) is expected to address this. • Convergence regime. The balanced model was trained for only 20 epochs; although the results are already competitive, a longer convergence regime has not been characterised, and the diminishing-returns region remains open. • Daylight and clear-weather dependence. The optical payload constrains the operational envelope to daylight and cloud-free scenes. Night and heavy-cloud operativity would require Synthetic Aperture Radar (SAR) or Infrared (IR) sensors, which are out of the scope of this work. • Generative background bias. The FLUX+LoRA pipeline is sensitive to the prompt and the trigger-token engineering; the model should be re-validated whenever the area of interest is changed, to prevent background bias from injecting artefacts into the synthetic distribution. These limitations are consistent with those reported in the broader edgeAI literature for space [13, 14, 12] and do not invalidate the core contribution of the paper, which is to demonstrate the feasibility of an integral workflow that jointly addresses on-board inference, generative data augmentation and SWAP-aware architecture selection for airborne surveillance from nanosatellites. 7. Conclusions and Future Work This paper has proposed and validated a workflow oriented to the deployment of computer-vision algorithms on board nanosatellites, showing the algorithmic and systemic feasibility of operating under the edge computing paradigm. The communication bottleneck of conventional bent-pipe architectures has been mitigated by moving inference on board a 6U CubeSat equipped with a Google Coral Edge TPU, so that only confirmed detection
35
metadata are downlinked. A SWAP-aware benchmark of three one-stage detectors has been carried out, and YOLO11n has been selected as the architecture that provides the best balance between precision and computational lightness, being the only candidate whose INT8-quantised weight (approximately 2.5 MB) fits the 8 MB on-die SRAM of the accelerator. The deficit of the minority class in the original HRPlanesV2 dataset has been resolved with a generative data-augmentation pipeline based on FLUX and LoRA: the synthetic, pseudo-labelled helicopter imagery has balanced the dataset to more than 53 000 instances and has lifted the helicopter F1 from 0.683 (classical augmentation alone) to 0.811, raising the global mAP@50 from 77.9 % to 82.2 %. Finally, the technical viability of the deployment on commercial hardware compatible with the SWAP budget of a CubeSat has been confirmed, with a projected on-orbit throughput of 25–30 FPS through SAHI inference of 1024 × 1024 patches. Despite these promising outcomes, the proposed approach still faces limitations, especially regarding the software-only validation of the on-board accelerator and the residual difficulty in detecting elongated infrastructure such as runways. The challenge lies in reconciling the dual objective of maximising detection precision and ensuring deployability under the strict SWAP constraints of a nanosatellite, a problem that intensifies as the diversity of target classes widens. Future work could explore several lines of research. First, the integration of SAR and IR sensors would enable 24/7 operativity regardless of meteorological conditions or the hour of the day. Second, moving from static image detection to on-board video tracking through algorithms such as ByteTrack or BoT-SORT, coupled with the already trained YOLO11n, would allow the estimation of kinematic parameters such as speed, heading and predicted trajectory. Third, the lightness of the nano architecture makes it exportable to Unmanned Aerial Vehicles (UAVs) as alternative host platforms, enabling real-time threat identification on board reconnaissance drones without downlinking the video to a command base. Fourth, the FLUX+LoRA pipeline could be redirected towards counter-camouflage training, forcing the detector to learn hidden patterns when an adversary attempts to conceal assets. Finally, the physical deployment of the compiled model on a space-grade accelerator — an Edge TPU or a Jetson Orin Nano — with measured power consumption and per-frame latency under vacuum and radiation, would provide the definitive validation of the concept and is the natural continuation of this work. 36
Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability The source code, trained weights, FLUX LoRA adapters and ComfyUI workflows supporting this study’s findings are publicly available in the following repository to ensure that the results are reproducible: https://github. com/Antonio23013/TFG-DETECCION-DE-OBJETOS.git. CRediT authorship contribution statement Antonio Delgado-Rosa: Writing – original draft, Visualisation, Validation, Software, Methodology, Data curation, Conceptualisation. David Muñoz-Valero: Conceptualization, Supervision, Methodology, Validation, Writing – review & editing. Enrique Adrian Villarrubia-Martin: Writing – review & editing, Data curation, Formal analysis. Juan MorenoGarcía: Conceptualization, Supervision, Formal analysis, Writing – review & editing, Funding acquisition. Acknowledgments This work was supported by grant PID2025-168152NB-C32 funded by MCIN/AEI/10.13039/501100011033, by ERDF A Way of Making Europe. It was completed when Enrique Adrian Villarrubia-Martin was a predoctoral fellow at the Universidad de Castilla-La Mancha funded by the European Social Fund Plus (ESF+). References [1] M. I. Skolnik, et al., Introduction to radar systems, volume 3, McGrawhill New York, 1980. [2] M. Strohmeier, M. Schäfer, V. Lenders, I. Martinovic, Realities and challenges of nextgen air traffic management: the case of ads-b, IEEE communications magazine 52 (2014) 111–118. 37
[3] M. N. Sweeting, Modern small satellites-changing the economics of space, Proceedings of the IEEE 106 (2018) 343–361. [4] A. Poghosyan, A. Golkar, Cubesat evolution: Analyzing cubesat capabilities for conducting science missions, Progress in Aerospace Sciences 88 (2017) 59–83. [5] G. Furano, G. Meoni, A. Dunne, D. Moloney, V. Ferlet-Cavrois, A. Tavoularis, J. Byrne, L. Buckley, M. Psarakis, K.-O. Voss, et al., Towards the use of artificial intelligence on the edge in space systems: Challenges and opportunities, IEEE Aerospace and Electronic Systems Magazine 35 (2020) 44–56. [6] A. Tchikladze, HRPlanesV2 computer vision model, https://universe.roboflow.com/alban-tchikladze-kxzqe/ hrplanesv2-7alnq, 2023. Último acceso: 26 de mayo de 2026. [7] B. F. Labs, Flux.1: Text-to-image generation models, 2024. URL: https://blackforestlabs.ai/. [8] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models, 2021. URL: https://arxiv.org/abs/2106.09685. arXiv:2106.09685. [9] European Space Agency, Deimos Space, Ciseres: Ai-powered satellites for rapid disaster response, https://www.esa.int/Applications/ Connectivity_and_Secure_Communications/Ciseres_AI-powered_ satellites_for_rapid_disaster_response, 2024. Último acceso: 26 de mayo de 2026. [10] S. Cass, Taking ai to the edge: Google’s tpu now comes in a makerfriendly package, IEEE Spectrum 56 (2019) 16–17. [11] R. Prokscha, M. Schneider, A. Höß, Efficient edge deployment demonstrated on yolov5 and coral edge tpu, in: Industrial Artificial Intelligence Technologies and Applications, River Publishers, 2023, pp. 141–155. [12] M. A. Chanoui, I. Khalil, M. Sbihi, Z. E. A. Alaoui Ismaili, Z. Guennoun, Trends and applications of on-board image processing for earth observation nanosatellites: a systematic review, International Journal of Aeronautical and Space Sciences 26 (2025) 1989–2019. 38
[13] G. Mystkowska, M. Monopoli, P. Nannipieri, L. Zulberti, D. M. Codinachs, L. Fanucci, Hardware platforms enabling edge ai for space applications: A critical review, IEEE Access (2025). [14] G. Lentaris, V. Leon, C. Sakos, D. Soudris, A. Tavoularis, A. Costantino, C. B. Polo, Performance and radiation testing of the coral tpu coprocessor for ai onboard satellites, in: 2023 European Data Handling & Data Processing Conference (EDHPC), IEEE, 2023, pp. 1–4. [15] S. C. Magalhães, F. N. dos Santos, P. Machado, A. P. Moreira, J. Dias, Benchmarking edge computing devices for grape bunches and trunks detection using accelerated object detection single shot multibox deep learning models, Engineering Applications of Artificial Intelligence 117 (2023) 105604. URL: https://doi.org/10.1016/j.engappai.2022. 105604. doi:10.1016/j.engappai.2022.105604. [16] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative adversarial networks, 2014. URL: https://arxiv.org/abs/1406.2661. arXiv:1406.2661. [17] C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of big data 6 (2019) 1–48. [18] S. Beery, G. Van Horn, P. Perona, Recognition in terra incognita, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 456–473. [19] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, Highresolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695. [20] P. N. Mueller, Attention-enhanced conditional-diffusion-based data synthesis for data augmentation in machine fault diagnosis, Engineering Applications of Artificial Intelligence 131 (2024) 107696. URL: https://doi.org/10.1016/j.engappai.2023.107696. doi:10. 1016/j.engappai.2023.107696. [21] N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, in: 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05), volume 1, Ieee, 2005, pp. 886–893. 39
[22] C. Cortes, V. Vapnik, Support-vector networks, Machine learning 20 (1995) 273–297. [23] S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. URL: https:// arxiv.org/abs/1506.01497. arXiv:1506.01497. [24] G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, L. Zhang, Dota: A large-scale dataset for object detection in aerial images, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3974–3983. [25] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779–788. [26] T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988. [27] Q. Liu, X. Xiang, Y. Wang, Z. Luo, F. Fang, Aircraft detection in remote sensing image based on corner clustering and deep learning, Engineering Applications of Artificial Intelligence 87 (2020) 103333. URL: https://doi.org/10.1016/j.engappai.2019.103333. doi:10. 1016/j.engappai.2019.103333. [28] D. İlmak, T. Bakirman, E. Sertel, Exploring you only look once v8 and v9 for efficient airplane detection in very high resolution remote sensing imagery, Engineering Applications of Artificial Intelligence 160 (2025) 111854. URL: https://doi.org/10.1016/j.engappai.2025.111854. doi:10.1016/j.engappai.2025.111854. [29] S. Liu, F. Shao, J. Xue, Recent advances of yolo-based object detection in remote sensing imagery: A comprehensive review, Engineering Applications of Artificial Intelligence 179 (2026) 115171. URL: https://doi.org/10.1016/j.engappai.2026.115171. doi:10. 1016/j.engappai.2026.115171. [30] X. Zhang, T. Lu, J. Wang, S. Fu, F. Gao, Small object detection by edge-aware neural network, Engineering Applications of Artificial 40
Intelligence 138 (2024) 109406. URL: https://doi.org/10.1016/j. engappai.2024.109406. doi:10.1016/j.engappai.2024.109406. [31] J. R. Wertz, D. F. Everett, J. J. Puschell, Space mission engineering: the new smad, (No Title) (2011). [32] B. Yost, S. Weston, State-of-the-art small spacecraft technology, Technical Report, 2024. [33] J. Puig-Suari, C. Turner, W. Ahlgren, Development of the standard cubesat deployer and a cubesat class picosatellite, in: 2001 IEEE aerospace conference proceedings (Cat. No. 01TH8542), volume 1, IEEE, 2001, pp. 1–347. [34] V. P. Katuntsev, Approaches to decompression safety support of eva for orbital and interplanetary missions, Acta Astronautica 66 (2010) 96–101. [35] T. Ma, M. Zhang, K. Chen, Z. Song, Y. Wu, X. Qian, Asymnvm: An efficient framework for implementing persistent data structures on asymmetric nvm architecture, in: Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, 2020, pp. 757–773. [36] S. S. Gao, Q. Luo, F. Zhu, Circularly polarized antennas, John Wiley & Sons, 2014. [37] J. Costantine, Y. Tawk, I. Maqueda, M. Sakovsky, G. Olson, S. Pellegrino, C. G. Christodoulou, Uhf deployable helical antennas for cubesats, IEEE Transactions on Antennas and Propagation 64 (2016) 3752–3759. [38] NVIDIA Corporation, NVIDIA T4 Tensor Core GPU Architecture, Whitepaper, NVIDIA, 2018. URL: https://www.nvidia. com/content/dam/en-zz/Solutions/Data-Center/tesla-t4/ t4-tensor-core-datasheet-951643.pdf. [39] E. Rapuano, G. Meoni, T. Pacini, G. Dinelli, G. Furano, G. Giuffrida, L. Fanucci, An fpga-based hardware accelerator for cnns inference on board satellites: Benchmarking with myriad 2-based solution for the cloudscout case study, Remote Sensing 13 (2021) 1518. 41
[40] Google LLC, Tensorflow models on the edge tpu, https://www.coral. ai/docs/edgetpu/models-intro, 2020. Último acceso: 07 de marzo de 2026. [41] Ultralytics, Ultralytics yolo export formats: Edge tpu, 2024. URL: https://docs.ultralytics.com/modes/export/, Último acceso: 07 de marzo de 2026. [42] TensorFlow, Post-training integer quantization, https://www. tensorflow.org/lite/performance/post_training_integer_ quant, 2023. TensorFlow Lite Documentation. Último acceso: 07 de marzo de 2026. [43] Simera Sense, Multiscape200, 2025. URL: https://simera-sense. com/products/multiscape200/, Último acceso: 18 de marzo de 2026. [44] Dragonfly Aerospace, Mantis imager, 2024. URL: //dragonflyaerospace.com/products/mantis/, Último 18 de marzo de 2026.
https: acceso:
[45] Tensor Tech, Attitude determination and control system for small satellites (>30 kg), 2025. URL: https://tensortech.co/product/detail/ adcs_for_over_30_kg, Último acceso: 18 de marzo de 2026. [46] R. Krenz, P. Sroka, M. Sybis, I. Zainutdinov, K. Wesołowski, A low-cost high-throughput c-band communication system for cubesats, Electronics 12 (2023). URL: https://www.mdpi.com/2079-9292/12/20/4374. doi:10.3390/electronics12204374. [47] Ericsson, Digital Airspace: The Sky - The Digital Frontier, Technical Report, Ericsson, 2020. URL: https://www.ericsson.com/en/ reports-and-papers/white-papers/digital-airspace, white paper, PDF. Último acceso: 18 de marzo de 2026. [48] T. Kuwahara, Introduction to cubesat command and data handling system, 2021. URL: https://www.unoosa.org/oosa/en/ourwork/psa/ hsti/kibocube.html, kiboCUBE Academy Lecture 10, PDF. Último acceso: 18 de marzo de 2026.
42
[49] A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. Parinov, M. Druzhinin, A. A. Kalinin, Albumentations: Fast and flexible image augmentation, Information 11 (2020) 125. doi:10.3390/info11020125. [50] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A. C. Berg, Ssd: Single shot multibox detector, in: European conference on computer vision, Springer, 2016, pp. 21–37. [51] G. Jocher, J. Qiu, Ultralytics yolo11, 2024. URL: https://github.com/ ultralytics/ultralytics. [52] Y. Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y. Liu, J. Chen, Detrs beat yolos on real-time object detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16965–16974. [53] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, S. Zagoruyko, End-to-end object detection with transformers, in: European conference on computer vision, Springer, 2020, pp. 213–229. [54] F. C. Akyon, S. O. Altinuc, A. Temizel, Slicing aided hyper inference and fine-tuning for small object detection, in: 2022 IEEE international conference on image processing (ICIP), IEEE, 2022, pp. 966–970.
43