IMAGIN-e: Onboard AI for Earth Observation

• IMAGIN-e Payload is a fully integrated ISS platform that demonstrates onboard AI inference using compressed Vision Transformer-based geospatial foundation models under strict resource constraints.
• The system employs a containerized, Linux-based environment with ARM Cortex-A72 cores running Docker and K3s to achieve inference results comparable to ground tests without dedicated accelerators.
• Key performance metrics include high detection accuracy for clouds and floods alongside a >16x model size reduction via distillation and FP16 quantization, paving the way for efficient spaceborne analytics.

The IMAGIN-e payload constitutes a fully integrated spaceborne platform deployed on the International Space Station (ISS) for on-orbit demonstration of geospatial foundation model (GeoFM) inference. Designed to validate the capacity for onboard AI-driven Earth observation (EO) under severe resource constraints, IMAGIN-e executes compact Vision Transformer-based (ViT) GeoFMs without hardware accelerators, demonstrating operational equivalence to ground validation in a real flight environment. The payload serves as a reference implementation for containerized, resource-efficient deep learning in space, merging state-of-the-art model compression, robust system architecture, and rigorous domain adaptation.

1. Hardware Configuration and Constraints

The compute backbone of IMAGIN-e is a 16-core ARM Cortex-A72 module (aarch64), typically operating at 1.2–1.5 GHz in comparable qualified boards. The system provides 16 GB DDR4 onboard memory, supporting 10 GB of flash storage for model binaries, container artifacts, and test data. No dedicated hardware accelerators (e.g., VPUs or GPUs) are made available; all inference is performed exclusively on the CPU. Sustained high-core utilization can approach the module’s thermal shutdown threshold, and in on-orbit trials, active core count was dynamically reduced to maintain safe margins. Peak per-core power consumption is estimated at 2–3 W under load (ground measurements), corresponding to module totals in the 30–40 W range, but the actual consumption depends on ISS environmental and utilization conditions.

Imaging instrumentation includes an RGB camera (operational but out of focus during demonstrations) and a HyperScout-M hyperspectral imager (0.45–0.95 µm, 50 bands, 75 m GSD), which was unavailable during the test due to a power fault. The overall enclosure fits the 3 U form factor (approximately 30 × 10 × 10 cm³, <25 kg), in line with external ISS payload module norms.

2. System Software and Operational Workflow

IMAGIN-e runs a Linux-based OS (aarch64) orchestrating containerized workloads using Docker and K3s, a lightweight Kubernetes distribution. Application containers are tailored for ARM Cortex-A, and all operations proceed within a fully offline environment (no external network access). System management leverages Microsoft’s Azure Orbital Space SDK for command orchestration.

GeoFM student models (e.g., 256-MAE-D) are exported to ONNX format and executed via ONNX Runtime on CPU, as no acceleration libraries (such as OpenVINO) are in use. The mission workflow employs ground-prepared image tiles (224 × 224 × 4 bands), uplinked to flash storage and processed by a Kubernetes job as a simple map-over-tiles operation. Each tile is normalized to [0,1], passed through the ViT encoder and its task head, and generates output masks or classification scores. Asynchronous or batched execution is not employed; per-tile inference is scheduled serially and outputs are downlinked as metadata.

3. Model Compression and Domain Transfer

Compression

The original Prithvi-EO-2.0-300M teacher model, sized at ∼1.2 GB (FP32), is distilled into a 256-MAE-D student of 73 MB (FP32), yielding a compression factor

$$r = \frac{1,200\,\text{MB}}{73\,\text{MB}} \approx 16.4.$$

Onboard quantization to FP16 further reduces storage to ∼36.5 MB, with minimal observed performance degradation (<0.2% accuracy loss).

Distillation and Adaptation

Both teacher and student employ a Masked Autoencoder (MAE) pipeline for dual-MAE knowledge distillation. For input patches $x$ and mask set $\mathcal{M}$, the loss minimized is

$$\mathcal{L}_{\mathrm{distill}} = \frac{1}{|\mathcal{M}|}\sum_{i\in\mathcal{M}} \|\hat x^s_i - \hat x^t_i\|_2^2,$$

where $\hat x^t_i$ and $\hat x^s_i$ denote teacher and student reconstructions. Post-pretraining, only the compact encoder is retained for inference.

Domain adaptation leverages Sentinel-2 imagery (source) and Kanyini cubes (target), extracting 4-band RGB+NIR tiles and generating automated proxy labels: water via NDWI (Otsu-thresholded), and clouds via HOT index (Otsu-thresholded). Three modes of adaptation are explored per task: (1) full fine-tuning (random ViT + head), (2) frozen GeoFM encoder + random head, and (3) frozen GeoFM + head pretrained on source. Training employs weighted cross-entropy with class weights

$$\{w_{\text{bg}}:w_{\text{fg}}\} = \begin{cases} 2:1 & \text{(clouds)}\ 1:4 & \text{(floods)} \end{cases}$$

and ∼1:1 sampling of contaminated/clear tiles to mitigate class imbalance.

4. On-Orbit Inference and Performance Benchmarks

Due to imaging subsystem issues, only pre-acquired, uplinked test sets were available for on-orbit evaluation. Results for onboard runs were numerically identical to those from ground engineering models (e.g., RPi 4):

Cloud Detection (Tile Classification, 792 tiles):

• Accuracy: 97.22%
• False-positive rate: 0.76%
• F1: 95.22%

Cloud Detection (Pixel Segmentation, 792 tiles):

• mIoU: 90.14%
• mF1: 94.80%
• Overall accuracy: 95.18%
• False-positive pixels: 1.50%

Flood Detection (Pixel Segmentation, 880 tiles):

• mIoU: 91.45%
• mF1: 95.46%
• Water-class IoU: 86.49%
• Water-class F1: 92.76%

Inference latency and throughput were not profiled in orbit; ground emulation shows ∼5.5 s per tile (∼180 tiles/hour), matched by on-orbit job timing.

5. Technical Lessons and Implementation Guidance

Thermal control emerged as a primary operational constraint; sustained CPU usage led to hardware-initiated shutdowns, necessitating core throttling. A plausible implication is that future CPU-only space payloads require integrated active thermal management or adaptive scheduling to maximize safe compute availability.

Model deployment viability on commodity ARM systems depends on aggressive compression. Factors of $r > 16$ reduction (via distillation and quantization) are critical for memory and storage compliance, and FP16 quantization yields substantial savings with negligible loss, especially when hardware acceleration is feasible.

Containerization (Docker plus K3s) offers robust cross-platform portability for both engineering and flight environments, though preprocessing toolchains (ONNX-to-IR, format compatibility) must target the specific onboard inference stack in advance.

Domain adaptation protocols based on a frozen ViT encoder with a retrained lightweight head are effective, with as little as 25% of full-label coverage in the target domain being sufficient for transfer stability. Pretrained heads further mitigate distributional shift. Joint constraints on uplink bandwidth, power, and thermal headroom dictate a holistic approach to pipeline design, covering model distillation, quantization, export, and container packaging on ground systems.

6. Significance and Future Implications for Onboard AI

IMAGIN-e demonstrates that compressed ViT-based GeoFMs, paired with lightweight adaptors, are operationally viable in stringent low-orbit environments, requiring no modifications from their ground-deployed state. The experiment underscores the practical pathway from large-scale pretraining to resource-constrained, flight-certified inference, enabling onboard scientific analysis within tight memory, storage, and energy budgets. These results establish a foundational blueprint for future EO missions seeking to embed advanced AI within extreme resource constraints, contingent on robust model compression and adaptive operational strategies (Du et al., 1 Dec 2025).