On-board satellite pipeline for detecting dark vessels (ships without AIS), using Liquid AI’s LFM2.5-VL as the decision engine at two stages: confirming NIR anomalies and classifying vessels in high-res imagery. Built in the final two hours of the Liquid AI × DPhi Space “AI in Space” hackathon.
Pipeline
| Stage | Filter | Result |
|---|---|---|
| 1. CloudFilter | SCL band cloud pixels | >70% cloud: discard |
| 2. AnomalyDetector | NIR threshold + VLM confirmation | No vessel: discard |
| 3. SuperResolution | Mapbox 0.5m/px + VLM classification | 128x128 crop + JSON transmitted |
Stack
- LFM2.5-VL (Liquid AI): 1.6B params, 2.6 GB quantized
- SimSat (DPhi Space): satellite simulator
- Sentinel-2: 10m multispectral imagery
- VRSBench: fine-tuning dataset (NeurIPS 2024)
- Modal: H100 for LoRA fine-tuning