Adaptive Onboard AI for EO Data Prioritization

Problem Statement

Earth observation satellites generate more imagery than they can transmit given limited onboard memory, power, and intermittent ground-station contact. Conventional 'downlink everything' approaches waste scarce bandwidth and delay access to time-critical information such as deforestation and wildfire detection.

Adaptive Scientific Value Function (ASVF) with Safety Gate

I worked on a adaptive scientific value function (asvf) with safety gate where:

• Onboard lightweight CNN (MobileNetV2) for real-time tile scoring
• Multi-signal ASVF combining detection confidence, event severity, and live resource state
• Mandatory safety gate ensuring critical alerts always transmit
• Skyfield-based orbital simulation for realistic communication windows
• Interactive dashboard for visualizing forest change, ASVF scores, and transmission decisions

My Contributions

• Designed the ASVF scoring framework fusing multi-signal environmental event detection
• Developed lightweight CNN pipeline on PyTorch for onboard imagery prioritization
• Built FastAPI backend for model inference, ASVF ranking, and resource simulation
• Implemented React + D3.js dashboard for live visualization of satellite resource states
• Integrated orbital simulation using Skyfield for realistic communication window modeling

Engineering Trade-offs

• Accepted reduced model capacity for onboard deployment versus cloud accuracy
• Chose transparent ASVF score over opaque black-box prioritization to ensure auditability
• Lowered update frequency to conserve satellite power while maintaining timely alerts

Impact