Orbital Edge Intelligence

The Shift to Edge AI

Traditional Earth Observation relies on a Store and Forward model: satellites capture massive amounts of raw data, wait for a ground station pass, and only then can analysis begin. This creates latency measured in hours or days.

SentianOrbit Edge AI flips this paradigm. By processing data onboard the satellite using low-power FPGAs, we convert gigabytes of raw pixels into kilobytes of actionable insights. Only the detections are sent down, enabling sub-minute response times.

• 99.9% Bandwidth Reduction: Sending metadata instead of raw SAR imagery.
• Real-time Alerting: Direct satellite-to-authority links for critical spill events.
• Autonomous Filtering: Ignoring empty ocean before transmission.

Computer Vision on SAR Data

Synthetic Aperture Radar sees through darkness and cloud cover, making it the only reliable sensor for 24/7 ocean monitoring. Our computer vision pipeline is built specifically for SAR characteristics.

• Spectral Normalization: Balancing high-dynamic range SAR signals and speckle noise.
• Sliding Window Inference: Processing 10,000x10,000 px scenes in optimised 416x416 tiles.
• Spatial Correlation: Mapping detections to GPS coordinates with meter-level precision.

YOLO26n: INT8 Quantized CNN

Our YOLO26n architecture is optimised for both speed and accuracy on SAR imagery. The full FP32 model is quantised to INT8 through a 500-chip calibration dataset, delivering 4x size reduction with less than 2% mAP loss.

• Convolutional Layers: Filters detect edges, textures and shapes of oil spills and vessels.
• INT8 Quantization: Weights compressed to 8-bit integers, fitting into FPGA BRAM.
• ONNX Runtime Web: The same INT8 model runs client-side in any modern browser.