Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation

Technical Summary: Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation

Problem Statement
The paper addresses two interconnected bottlenecks hindering airborne surveillance from Low Earth Orbit (LEO) nanosatellites. First, the "bent-pipe" operational paradigm, inherited from early space missions, requires satellites to downlink terabytes of raw imagery for ground processing. This approach is constrained by the limited downlink budget of nanosatellites (typically 10–30 W power budgets and low-profile antennas), creating a latency of hours or days that prevents real-time decision-making. Second, the training data for aircraft detection is scarce and severely class-imbalanced. Open datasets like HRPlanesV2 contain a disproportionate number of military and civil aircraft compared to minority classes like helicopters (a ~4:1 ratio), preventing standard detectors from learning robust representations of rare targets. Existing literature typically addresses these issues in isolation: either by running heavy models on the ground (ignoring downlink limits) or by training on imbalanced data (ignoring minority class performance), often without considering the strict Size, Weight, and Power (SWaP) constraints of space hardware.

Methodology
The authors propose a unified workflow that integrates on-board inference with generative data augmentation, designed specifically for a 6U CubeSat platform equipped with a Google Coral Edge Tensor Processing Unit (TPU).

1. Hardware-Constrained Architecture Selection:
   The methodology adopts a "hardware-first" approach. The target platform is a 6U CubeSat with an 8 MB on-die Static Random-Access Memory (SRAM) limit on the Edge TPU. The authors evaluated three one-stage detectors: SSD MobileNet V3, RT-DETR-L, and YOLO11n.

   • SSD MobileNet V3 was discarded due to poor detection accuracy (mAP@50 of 44.98%).
   • RT-DETR-L was rejected because its memory footprint (188.85 MB) vastly exceeded the 8 MB SRAM limit, even after INT8 quantization.
   • YOLO11n was selected as the optimal candidate. After INT8 quantization, its weight (~2.5 MB) fits within the SRAM budget, allowing the entire inference path to run on-chip without CPU offloading.

2. Generative Data Augmentation Pipeline:
   To resolve the class imbalance, specifically for the minority helicopter class, the authors developed a pipeline using the FLUX diffusion model and Low-Rank Adaptation (LoRA).

   • A custom LoRA adapter was trained on the limited real helicopter samples using a specific trigger token (sathelo).
   • The model generated synthetic helicopter imagery across four operational environments (industrial, rural, coastal, arid).
   • These synthetic images were automatically annotated (pseudo-labeled) by an intermediate YOLO detector and merged with classically augmented samples (geometric and photometric transformations) to create a balanced dataset of over 53,000 instances.

3. On-Board Concept of Operations (ConOps):
   The system operates under a "total-decision autonomy" model. The satellite activates only over pre-programmed coordinates. The on-board AI processes the video stream in real-time using Slicing Aided Hyper Inference (SAHI) to handle 1024×1024 patches. Frames without detections are discarded; only metadata (coordinates, confidence, timestamp) and compressed patches of positive detections are downlinked via an S-band link, drastically reducing bandwidth usage compared to raw image transmission.

Key Contributions

• Hardware-First Methodology: A design process where the physical constraints of the 6U CubeSat and the 8 MB SRAM limit of the Edge TPU dictate the selection of the AI architecture, ensuring deployability.
• Generative Augmentation for Minority Classes: A novel pipeline combining FLUX and LoRA to synthesize minority-class imagery (helicopters), which is then pseudo-labeled and integrated to balance the training dataset.
• SWaP-Aware Benchmark: A comparative analysis of SSD MobileNet V3, RT-DETR-L, and YOLO11n under strict space-grade constraints, identifying YOLO11n as the only viable architecture for this specific memory budget.
• Open-Source Reproducibility: Release of source code, trained weights, FLUX LoRA adapters, and ComfyUI workflows to allow community verification and expansion.

Results

• Detection Performance: The balanced dataset achieved a global mean Average Precision (mAP@50) of 82.2%, an increase from 77.9% in the unbalanced baseline.
• Minority Class Improvement: The F1-score for the minority helicopter class improved significantly from 0.683 (using only classical augmentation) to 0.811 (using generative augmentation), approaching the performance of majority classes.
• Deployment Feasibility: The quantized YOLO11n model fits entirely within the 8 MB SRAM. Based on throughput-to-power scaling, the authors project an on-orbit inference rate of 25–30 Frames Per Second (FPS) at 1024×1024 resolution.
• Extended Capabilities: A proof-of-concept extended to 22 specific aircraft classes (including F-16, B-52, etc.) achieved >95% accuracy for most classes, though runway detection remained lower (37%) due to confusion with linear background structures.

Significance and Claims
The paper claims to demonstrate the feasibility of a decision-support tool for real-time, autonomous airborne surveillance from nanosatellites. By shifting from a passive data collector to an active inference node, the proposed workflow mitigates the downlink bottleneck, reducing decision latency from hours/days to fractions of a second. The authors assert that this approach is energetically sustainable, as the power cost of the Edge TPU is offset by the massive reduction in radio-frequency transmission requirements.

The paper maintains a modest stance regarding its validation, acknowledging that all performance metrics (FPS, latency) are projections based on scaling ground-based GPU measurements to the Edge TPU's throughput-to-power ratio. It explicitly states that no Hardware-in-the-Loop tests or radiation hardness verification (Single Event Effects/Total Ionizing Dose) have been conducted on flight-grade hardware yet. Furthermore, the authors note limitations in detecting elongated infrastructure (runways) and the current dependence on daylight and clear weather, identifying these as areas for future work involving SAR/IR sensors and instance segmentation.