Towards Autonomous Aircraft Surveillance from Nanos...

Towards Autonomous Aircraft Surveillance from Nanos... | AI Research

Key Takeaways

Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation proposes a method to enable nanosate...
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.

Paper AbstractExpand

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 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.

Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation proposes a method to enable nanosatellites to identify aircraft in real-time without needing to transmit massive amounts of raw imagery to ground stations. By combining on-board AI processing with synthetic data generation, the researchers aim to solve the dual problems of limited satellite bandwidth and the scarcity of high-quality training data for rare aircraft classes.

Addressing the Data Bottleneck

Current satellite surveillance typically follows a "bent-pipe" architecture, where satellites act as passive collectors that transmit all raw imagery to the ground. This creates a bottleneck because the volume of data generated by high-resolution sensors far exceeds the capacity of satellite radio-frequency links. The authors propose replacing this with an autonomous, on-board processing workflow. By performing inference directly on a 6U CubeSat, the satellite can transmit only lightweight metadata—such as target coordinates and timestamps—instead of gigabytes of raw pixels.

Generative Data Augmentation

A major obstacle in training accurate detection models for space is the class imbalance in existing datasets, where common aircraft significantly outnumber rare ones like helicopters. To address this, the researchers developed a generative pipeline using FLUX and Low-Rank Adaptation (LoRA). They fine-tuned a diffusion model to create synthetic images of minority-class aircraft. These synthetic samples were then automatically labeled by an intermediate detector and combined with traditional data augmentation to create a balanced training set. This process improved the F1 score for the minority class from 0.683 to 0.811.

Hardware-First Deployment

The researchers employed a "hardware-first" methodology, ensuring their AI models fit the physical constraints of a 6U CubeSat before training. They selected the Google Coral Edge TPU as the on-board accelerator, which requires models to be quantized to 8-bit integers (INT8) and fit within an 8 MB on-die Static Random-Access Memory (SRAM) limit. Among the architectures tested—SSD MobileNet V3, YOLO11n, and RT-DETR-L—the team identified YOLO11n as the best balance between precision and latency.

Performance and Results

The proposed workflow achieved a global mean average precision (mAP) of 82.2%, up from 77.9% using standard methods. When integrated with Slicing Aided Hyper Inference (SAHI) to process 1024 × 1024 pixel patches, the system is projected to perform at 25–30 frames per second on orbit. The authors note that while this approach improves autonomous surveillance capabilities, it is constrained by the specific power and memory limitations of the chosen hardware, which necessitates the use of INT8 quantization and careful model selection to avoid offloading computations to the host CPU.