AI-DRIVEN ONBOARD SATELLITE PROCESSING: A PARADIGM SHIFT IN EARTH OBSERVATION

Abstract

The surge of Earth Observation (EO) satellites has resulted in unprecedented data volumes, frequently exceeding the downlink and processing capabilities of ground stations. Traditional workflows—capture, downlink, process—are considered inefficient for time-critical applications such as disaster response, maritime surveillance, and climate monitoring. In this context, AI-driven onboard satellite processing is explored, where artificial intelligence (AI) and GPU/AI accelerator hardware (e.g., NVIDIA Jetson, Intel Myriad) are deployed directly on satellites to enable real-time, in-orbit analytics. Building on missions such as ESA’s PhiSat-1 (2020, Myriad-2 VPU), PhiSat-2 (2024, Intel Movidius VPU) and Orbital Sidekick (2021, Jetson TX2), concrete demonstrations of onboard AI for cloud filtering, anomaly detection, hyperspectral leak monitoring, and object recognition are analyzed. Results have shown up to a 95% reduction in downlink data volume and millisecond-scale inference times, demonstrating that onboard computing significantly enhances satellite efficiency.     The role of parallel computing architectures (GPUs, VPUs, FPGAs) is further examined, performance benchmarks are discussed, and future directions for autonomous, adaptive, and swarm-enabled satellite constellations are forecast. Findings highlight that shifting computation from ground to orbit fundamentally changes EO economics, enabling a proactive, responsive, and intelligent space infrastructure.